Traffic Optimization Using Machine Learning

Making intersections smarter to reduce driver frustration
By Joe Kessler – Principal Software Developer

🚦 Overview

This project implements a machine learning-driven traffic simulation for optimizing signal timing in a small city grid. The goal: reduce travel times and improve traffic flow using a Convolutional Neural Network (CNN) and real-time image data from simulated traffic cameras.

🧠 Motivation

• Traffic lights often use simplistic or rigid logic.
• Drivers get stuck at red lights unnecessarily.
• Poor traffic flow increases fuel usage and stress.

Can we do better?
Yes—by simulating intersections that “see” traffic via cameras and respond intelligently.

---

🏙️ Simulation Environment

• 3×3 grid of intersections (9 total)
• Each intersection equipped with 4 cameras (North, South, East, West)
• Real-time image processing to detect traffic and make decisions
• Cars behave like agents with basic GPS routing and 40 MPH max speed

---

🧩 Components

Component	Language/Tech	Description
Simulator	Python	Core traffic simulation environment
CNN Model	Python (Keras/TensorFlow)	Detects cars and zones from camera images
Training Data Generator	C#/WinForms	Creates labeled traffic image data for training
Traffic Visualizer	C#/WinForms	Visual display of traffic activity
CameraWS	C#/ASP.NET	Simulated camera feed for intersections

---

📷 Neural Network: CNN for Traffic Vision

The CNN processes images from simulated cameras to:

• Count vehicles in each lane
• Estimate vehicle distance in three zones: Far, Middle, Near
• Output 9 values per image (3 lanes × 3 distance zones)

Sample Model Code

model = Sequential()
model.add(Conv2D(32, (5, 5), activation='relu', input_shape=imageShape))
model.add(MaxPooling2D((2, 2)))
model.add(Conv2D(32, (5, 5), activation='relu'))
model.add(MaxPooling2D((2, 2)))
model.add(Flatten())
model.add(Dense(100, activation='relu'))
model.add(Dense(100, activation='relu'))
model.add(Dense(9, activation='sigmoid'))

model.compile(optimizer='rmsprop', loss='mean_squared_error', metrics=['accuracy'])
model.fit(trainX, trainY, batch_size=1, epochs=250, verbose=1)

---

🏃 Running the Simulation

Once the CNN is trained, it integrates seamlessly into the simulator:

1. Capture images from all 4 directions via CameraWS.
2. Predict vehicle presence using CNN.
3. Use simple logic to decide if a light should switch.

Example (simplified):

images = capture_image(myconstants.camera_ws_url)
for direction in images:
    prediction = model_cnn.predict(images[direction])
    traffic_scenario = TrafficScenario(self, prediction, road_lanes, self.global_ctx)
    predictions[direction] = traffic_scenario

---

📊 Results

• Same car throughput
• ~50% fewer signal changes
• ~14% lower congestion
• ~11% lower travel times
• Drivers maintain momentum better—fewer “stop-and-go” scenarios

---

🛠️ Hardware Used

• Intel Core i7-12700K
• 128GB DDR4-3600 RAM
• NVIDIA RTX 3090 Ti (24GB VRAM)

---

🙋 Questions?

This was originally a video presentation!
For questions/comments, reach out to Joe Kessler.
There’s no complaint box… yet.